Event-Type Identification in Power Grids Using a Spectral Correlation Function-Aided Convolutional Neural Network

Abstract

Rapid and accurate identification of events in power grids is critical to ensuring system reliability and security. This study introduces a novel event-type identification method, utilizing a Spectral Correlation Function (SCF)-aided Convolutional Neural Network (CNN). The proposed method employs a six-stage cascaded structure consisting of: (1) data collection, (2) clipping, (3) augmentation, (4) feature extraction (FE), (5) training, and (6) testing. Real-world power grid signals sourced from the Grid Event Signature Library are used for both training and testing. To improve robustness, additive white Gaussian noise (AWGN) is introduced at various signal-to-noise ratio (SNR) levels to augment the dataset. The SCF-based FE method captures distinctive event-type characteristics by exploiting the spectral correlation of signals, allowing the CNN architecture to effectively learn and generalize event patterns. The proposed method is benchmarked against seven conventional techniques, using real-world power grid signals representing four distinct event types: blown fuse, line switching, low amplitude arcing, and transformer energization. Key performance metrics-prediction accuracy, mean absolute error (MAE), precision, recall, F1-score, and confusion matrix—are employed to evaluate the performance. Results demonstrate that the SCF-CNN method outperforms traditional approaches across all metrics and SNR levels, achieving over 99% prediction accuracy and nearly zero error for SNR values above 6 dB. This signifies its efficacy in reliable event-type identification for power grid applications.

Keywords

• Data augmentation
• event and fault classification
• grid event signature library (GESL)
• low amplitude arcing
• machine learning
• spectral correlation function (SCF)